Wearable Pressure Sensor Array for Health Monitoring

Matti Kaisti1, Joni Leppanen2, Olli Lahdenoja1, Pekka Kostiainen2,Mikko Pankaala1, Ulf Meriheina2,Tero Koivisto1

1 Department of Future Technologies, University of Turku, 20500 Turku, Finland2 Murata Electronics OY, Vantaa, Finland

Abstract

Measuring the arterial waveform in real-time usingwearable devices mounted directly on skin holds promisein assessing cardiovascular health status and detectingan early onset of cardiovascular disease. We report theuse of modern high performance MEMS pressure sensorsfor wearable health monitoring. The low-cost sensor el-ements were incorporated onto a flexible wristband forradial artery pulse measurement. These sensor elementswere configured as an array and attached to a wristband.The device operation was tested on 13 healthy subjects andfrom each subject we successfully derived the average ar-terial waveform, located the diastolic and systolic peakstogether with Dicrotic notch and calculated the heart rate.In the future, the MEMS pressure sensors might be em-ployed for mobile and remote cardiovascular health moni-toring.

1. Introduction

The first arterial waveform measurements were con-ducted already in the late 19th century. A device calledsphygmograph was used to measure blood pressure. Thismechanical device was soon replaced with a cuff basedmethod that is still prominently used today and is one ofthe most established diagnostic tools for blood pressuremonitoring. However, it is designed to record only twopressure values, the systolic and diastolic pressure. It cannot be used for continuous measurement although the latestcommercial devices are able to detect irregular beats (e.g.atrial fibrillation) and morning hypertension. By recordingthe arterial waveform a wealth of information about car-diovascular health and aging can be extracted. The arterialwaveform analysis allows the calculation of many parame-ters related to health. A modern technique for this purposeis called applanation tonometry where a sensor is pressedagainst an artery by a trained physician, and the pressureof the blood circulation is registered. These systems, how-ever, are not designed for wearable nor continuous use.

A related techniques for blood pressure and wave-

form sensing include the use infrared photoplethysmogram(PPG) placed e.g. in the finger with an inflatable cuff. Theblood volume changes can then be recorded with the helpof pneumatic valve [1] and controller electronics. Bloodpressure can also be computed from the pulse transmittime (PTT) when several sensing points are utilized (seee.g. [2]). One alternative to PPG for health status moni-toring is pressure sensing and the recent developments in-clude the following; a pressure sensor elements have beencreated using polymer transistors [3], gold nanowires [4]and piezoelectric materials [5] for measuring blood pulsewaveform. For health parameter extraction an automaticdelineation of arterial blood pressure waveform was devel-oped [6] and a single fiducial point method for long termblood pressure monitoring has been conceptualized [7].

We concentrate on the waveform analysis of a non-invasive wearable and continuous-time MEMS arterialpressure sensing device. The use of modern MEMS sensortechnologies have high performance in a compact packagewith low power consumption. These sensors can be seam-lessly integrated into modern electronic devices. Since themeasurement procedure is sensitive to the exact placementof the sensor to an artery, it is preferable to have severalsensors arranged into a grid and use an automated selec-tion of the best MEMS sensor from the multiple of sensingelements. With each recording a high quality signal wasobtained from the sensor array and subsequently the heartrate was computed and arterial waveform extracted indi-cating the sensors potential for wearable health monitoringin the future.

2. Methods

Sensor operation principle: Murata’s capacitive pressureelement (SCB10H) consists of two silicon wafers and oneglass wafer. One silicon wafer is the sensors diaphragmthat bends with external pressure. This pressure is propor-tional to the amount of bending and therefore the change indistance between two electrodes alters the sensors capaci-tance. The element is an absolute pressure sensor and thetypical dynamic range is 4 pF or 50 % of the base capac-

Computing in Cardiology 2017; VOL 44 Page 1 ISSN: 2325-887X DOI:10.22489/CinC.2017.143-140

0 5 10 15

0 5 10 15

0 5 10 15

0 0.5 1 1.5 2 2.5 3Time (s)

A

B

C

D

Figure 1: A) Raw signal traces from the array. B) Theselected signal after band-pass filtering. C) Detected peaks(red), corresponding valleys (blue) and adaptive threshold(dashed line) which was used to determine whether eachpeak candidate is a true peak or a noise peak. D) A zoomof C) with only found peaks indicated (red).

itance. It is only limited by the allowed maximum non-linearity associated with the hyperbolic relationship be-tween electrode distance and capacitance value. This cre-ates a significantly high dynamic range and gives a uniquecombination of high resolution and low current consump-tion. Murata’s capacitive design in a single crystal siliconalso results in minimized material stress, maximized over-load performance and has practically no plastic deforma-tion. When the element is pressed against an artery theobtained pressure value is comprised from counterforce ofthe hold down pressure as well as from the transmural pres-sure from the blood vessel [8]. The measured pressure isa sum of these components. The hold down pressure in anideal case remains constant. Movement and arm tensioncan create significant changes in this pressure componentcausing difficulty in detecting the pulse waveform.Signal acquisition system: The pressure sensor hous-ing was covered with gel for coupling the blood pressurepulse signal to the sensor element. The change in thesensor capacitance was read with Silicon labs EFM32, a32-bit ARM Cortex-M3 core based microcontroller, us-ing the Acam PCap04Z capacitance-to-digital converter. APython application was used to acquire the signal from themicrocontroller to the PC via RS232 serial interface.Data collection: Three pressure sensor elements(SCB10H Murata) were assembled on a flexible wristband.The band was wrapped around the left wrist and positionedroughly over the radial artery. The output was digitized and

send to PC for post processing. Data was collected from13 healthy subjects between 25 and 53 years of age. Allsubjects were in a seated position with the arm in a relaxedposition. The sensor array and the wristband placement areshown in Figure 2 b).

3. Algorithm

Preprocessing and sensor selection: The preprocessingof each sensor element signal include decimating it from685 Hz to 200 Hz signal, removing the mean values ofthe signal and clipping spurious glitches against predeter-mined threshold. From the preprocessed signal we com-pute all local peaks and valleys that have atleast 0.5 s sep-aration. The median value is computed for the peaks andvalleys from which a single peak-to-peak value is com-puted. This value is divided by the standard deviation ofthe high-pass (fc = 50 Hz) filtered version of the signal.The signal with the highest value is considered to have thehighest signal-to-noise ratio and is selected for further pro-cessing. The selected signal was band-pass filtered usinga third order Butterworth zero-phase IIR filter with cut-offfrequencies of 1 Hz and 20 Hz. In addition to high fre-quency noise the filter removes bias and partly the baselinewandering.Peak detection: After filtering the selected signal thepeaks are detected using an automatic multiscale-basedpeak detection (AMPD) algorithm [9]. The algorithm con-structs a matrix consisting of scale-dependent local max-imas. The details for the peak detection are provided in[9]. The found local maximas are considered as peak can-didates and a search-back routine is initiated that deter-mines whether the candidate is either a noise or a true peak.This step is based on adaptive thresholding. For each peakcandidate amplitude Pa the threshold is determined by aweighted average of Pa and current threshold (THi) byTHi+1 = 1/3× Pa + 2/3× THi. Subsequently the THis updated if the following condition of a true peak is satis-

(a) (b)

Figure 2: a) MEMS pressure sensor array configurationon small PCB with interface circuitry. b) The wearablewristband placement during measurement.

Page 2

fied, Pa > 0.7× THi+1. We do not initiate a search-backif a peak is clearly missed as the purpose of the algorithm isto robustly find correct peaks for waveform analysis ratherthan accurate heart rate estimation although the algorithmfinds peaks from the test data set with high performance.Average waveform: The peaks were used as fiducialpoints for waveform ensemble averaging. The averagetime difference of preceding valleys before each peak wasused to compute the window length for the waveform.Each obtained waveform (one for each found peak) werealigned by matching the peaks. Waveforms deviating morethan two standard deviations from the mean were dis-carded. The remaining waveforms were ensemble aver-aged and the result is the final pulse waveform.Waveform analysis: From the obtained waveform the sig-nal was spline interpolated to 1000 Hz and normalizedto have a maximum of one. The dicrotic notch was de-termined by fitting a first order polynomial with a leastsquares mean from the the diastolic peak onward and sub-stracting this line from the original signal. This minimumpoint serves as another fiducial point that allows findingthe dicrotic notch from the original waveform. This isachieved by computing the minima from a narrow win-dow surrounding the fiducial point. Furthermore, the 1stand 2nd order derivatives are computed and the zero cross-ings and local maxima and minima are computed fromthem, respectively. The dicrotic notch, systolic and dias-tolic peaks can also be computed from the signal deriva-tives as indicated in Figure 3. Combining these methodsallows more robust parameter estimation with inter-subjectvariability.

4. Results and Discussion

Sensor selection: The array configuration provides morerobust sensing and easier sensor placement for obtainingpulse waveform from the wrist compared to single ele-ment configurations. This, however, adds the requirementto either combine or select the best element from the array.We developed an sensor selection method. In the 13 mea-surements the quality was evaluated via an expert opinion.The expert considers the following quality metrics: abil-ity to detect peaks clearly, the shape of the pulses and theamount of high frequency noise. In 85 % of the cases thealgorithm and expert opinion match. In the remaining twocases the algorithm chooses an element that is consideredonly slightly sub-optimal. In these cases two elements pro-vide a high quality signal. In all cases a clear signal wave-form and heart rate can be obtained.Heart-rate detection: The peak detection algorithm wasevaluated by comparing the detected peaks to an ex-pert annotation. The detection performance of heartbeatswere evaluated by calculating the positive predictive value(PPV), and the true positive rate (TPR). The results are col-

0

0.2

0.4

0.6

0.8

1

−0.2

0

0.2

0.4

0.6

0.8

0.2 0.4 0.6 0.8−1

−0.5

0

0.5

1

Diastolic Peak

Systolic Peak

DicroticNotch

1st derivative

2nd derivative

Time (s)

Figure 3: a) Extracted ensemble averaged waveform. b)The 1st derivative of the waveform. c) 2nd derivative ofthe waveform. Systolic and diastolic peaks and dicroticnotch are indicated in the figure with (red) circles.

lected to Table 1 where the high performance of the detec-tion is evident. In most cases, the high signal quality sim-plifies the actual detection significantly. These rates are,however, expected to be clearly worse in a less controlledenvironment and in the presence of motion artefacts.Radial artery pulse measurement: One of the sensorlimitations arise from the hand movement and sensor po-sitioning. Any tension in the arm affects the sensor out-put via changes of pressure by either conformal changes

Page 3

Table 1: Performance metrics of the beat-to-beat detection.HR, TPR and PPV stand for heart rate, true peak rate andpositive predictive value respectively. Bottom line showsthe mean of each metric.

ID HR (bpm) TPR (%) PPV (%)1 90.3 96.9 99.52 72.0 100 1003 67.5 100 1004 91.8 92.7 1005 57.5 99.2 1006 93.4 100 1007 61.6 100 1008 107.6 100 1009 96.9 100 10010 69.9 100 10011 60.8 100 10012 60.2 99.2 10013 74.3 100 100x 77.2 99.1 100

under the sensing area of changes via the arm diametercausing strap to be pressed against the arm with differentforce. The operation in the current configuration is mainlylimited to still operation. This, however, only partly lim-its the usability as the waveform can be obtained usinga very short measurement time. An example figures forthe ensemble averaged waveform and its 1st and 2nd orderderivatives is presented in Figure 3. These allows anno-tating and deriving the determination of important wave-form parameters. We demonstrate the extraction of thesystolic and diastolic peaks and dicrotic notch. The rela-tion of these parameters to cardiovascular health has beenreviewed [10].

5. Conclusion

The presented wearable device and the accompanyingalgorithm indicate that MEMS pressure sensors could beused for future wearable skin-mounted sensors for long-term and unobtrusive monitoring. The sensor operationrequires little or no training and can be fabricated withlow-cost. The high performance operation of the MEMSdevices provide side-effect free daily operation and theplastic cover ensures that there is no skin irritation. Thedownside of the sensing configuration is its sensitivity tohand movement and muscle tension which cause signifi-cant amplitude variations to the signal. Thus, the systemis probably limited to applications where it suffices to takecontinuous point in time measurements. The applicabil-ity of the sensor array and accompanying algorithm wasdemonstrated via sensor element selection after which theheart beats could be robustly detected together with a clean

waveform. Future work includes investigating the abil-ity to extract artefacts from the signal, detection of atrialfibrillation and a comparison study with invasive pressuresensors.

References

[1] Parati G, Casadei R, Groppelli A, Di Rienzo M, Mancia G.Comparison of finger and intra-arterial blood pressure mon-itoring at rest and during laboratory testing. Hypertension1989;13(6 Pt 1):647–655.

[2] Gesche H, Grosskurth D, Kuchler G, Patzak A. Continuousblood pressure measurement by using the pulse transit time:comparison to a cuff-based method. European journal ofapplied physiology 2012;112(1):309–315.

[3] Schwartz G, Tee BCK, Mei J, Appleton AL, Kim DH, WangH, Bao Z. Flexible polymer transistors with high pressuresensitivity for application in electronic skin and health mon-itoring. Nature Communications 2013;.

[4] Gong S, Schwalb W, Wang Y, Chen Y, Tang Y, Si J, Shirin-zadeh B, Cheng W. A wearable and highly sensitive pres-sure sensor with ultrathin gold nanowires. Nature Commu-nications 2014;.

[5] Dagdeviren C, Su Y, Joe P, Yona R, Liu Y, Kim YS, HuangY, Damadoran AR, Xia J, Martin LW, Huang Y, RogersJA. Conformable amplified lead zirconate titanate sensorswith enhanced piezoelectric response for cutaneous pres-sure monitoring. Nature Communications 2014;.

[6] Li BN, Dong MC, Vai MI. On an automatic delineator forarterial blood pressure waveforms. Biomedical Signal Pro-cessing and Control 2010;5(1):76–81.

[7] Addison PS. Slope transit time (stt): A pulse transittime proxy requiring only a single signal fiducial point.IEEE Transactions on Biomedical Engineering Nov 2016;63(11):2441–2444.

[8] Zuo W, Wang P, Zhang D. Comparison of three differenttypes of wrist pulse signals by their physical meanings anddiagnosis performance. IEEE Journal of Biomedical andHealth Informatics Jan 2016;20(1):119–127. ISSN 2168-2194.

[9] Scholkmann F, Boss J, Wolf M. An efficient algorithmfor automatic peak detection in noisy periodic and quasi-periodic signals. Algorithms 2012;5(4):588–603.

[10] Elgendi M. On the analysis of fingertip photoplethysmo-gram signals. Current Cardiology Reviews 2012;.

Address for correspondence:

Matti KaistiDepartment of Future Technologies, University of TurkuVesilinnantie 5, 20500 Turku, Finlandmkaist@utu.fi

Page 4